Research map / Image generation / Image restoration
Diff-Shadow: Global-guided Diffusion Model for Shadow Removal
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Diff-Shadow combines a patch-based local diffusion UNet with a parallel low-resolution global UNet, linked by a Reweight Cross Attention module and a Global-guided Sampling Strategy, for shadow removal on ISTD, ISTD+ and SRD.
Problem
Shadow removal with diffusion models suffers from illumination inconsistency between shadow and non-shadow regions, and patch-based diffusion produces boundary artifacts because it lacks global context.
Approach
A parallel dual-UNet diffusion architecture in which a global low-resolution branch supplies non-shadow context to the local patch branch via mask-reweighted cross attention, plus a sampling strategy that fuses overlapping patch noise predictions using brightness consistency with the global output and the non-shadow ratio of each patch.
Key results
- On ISTD, Diff-Shadow reaches 33.69 dB PSNR overall versus 32.33 dB for ShadowDiffusion and 32.21 dB for ShadowFormer.
- On ISTD+, overall PSNR is 36.47 dB, above ShadowDiffusion's 35.68 dB.
- On SRD, overall PSNR is 34.93 dB versus 34.73 dB for ShadowDiffusion; ablations show RCA and GSS each reduce RMSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the Diff-Shadow global-guided diffusion method | Shadow removal · ISTD (All Image) | PSNR: 32.33 dB to 33.69 dB | Helped |
| Add the Diff-Shadow global-guided diffusion method | Shadow removal · SRD (All Image) | PSNR: 32.90 dB to 34.93 dB | Helped |
| Remove the global branch (ablation) | Diff-Shadow · SRD | RMSE (All): 2.52 -> 3.36 | Hurt |
| Concatenate patch and global images at the network input instead of using parallel UNets (ablation) | Diff-Shadow · SRD | RMSE (All): 2.52 -> 3.04 | Hurt |
| Remove RCA from the first layer (ablation) | Diff-Shadow · SRD | RMSE (All): 2.52 -> 2.73 | Hurt |
| Remove GSS and patch overlap (ablation) | Diff-Shadow · SRD | RMSE (All): 2.52 -> 3.81 | Hurt |
| Set patch fusion weights to W^d=1 (ablation) | Diff-Shadow GSS · SRD | RMSE (All): 2.52 -> 3.27 | Hurt |
| Use only brightness-based score S_bri for fusion weights (ablation) | Diff-Shadow GSS · SRD | RMSE (All): 2.52 -> 2.96 | Hurt |
| Use RCA in the first layer and remove CA (ablation) | Diff-Shadow · ISTD+ | PSNR: 36.47 -> 35.31 | Hurt |
| Remove RCA and use CA in the mid-layer (ablation) | Diff-Shadow · ISTD+ | PSNR: 36.47 -> 35.49 | Hurt |
| Move RCA from the first to the second layer (ablation) | Diff-Shadow · ISTD+ | PSNR: 36.47 -> 36.27 | Hurt |
| Move RCA to the third layer (ablation) | Diff-Shadow · ISTD+ | PSNR: 36.47 -> 35.93 | Hurt |
Surprising findings
- Adding RCA to both shallow and intermediate layers degraded performance, attributed to limited mask size.
Limitations
- Relatively long inference time compared with end-to-end restoration networks and simple-UNet diffusion methods.
- Requires accurate shadow masks; mask quality directly determines removal results.
- Failure cases on complex scenes with large shadow areas (CUHK-Shadow).
Experimental setup
Datasets: ISTD ISTD+ SRD CUHK-Shadow
Benchmarks: ISTD ISTD+ SRD
Metrics: PSNR SSIM RMSE (LAB)
Baselines: ShadowFormer ShadowDiffusion DHAN BMNet DC-ShadowNet ST-CGAN DeshadowNet Fu et al. Zhu et al. SG-ShadowNet Guo et al. CRFormer
Builds on: ShadowDiffusion conditional diffusion model WeatherDiffusion ShadowFormer
Closest papers by meaning
- Detail-Preserving Latent Diffusion for Stable Shadow Removal (CVPR 2025 · similarity 0.90)
- Polarization Guided Mask-Free Shadow Removal (AAAI 2025 · similarity 0.88)
- HomoFormer: Homogenized Transformer for Image Shadow Removal (CVPR 2024 · similarity 0.87)
- Relay Diffusion: Unifying diffusion process across resolutions for image synthesis (ICLR 2024 · similarity 0.87)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.87)
- Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.